Inspiration
A microbusiness scans its paperwork to a NAS because that is where the scanner sends it. Years later there are thousands of invoices, receipts, purchase orders and payment records sitting in folders, and at tax time the questions are never about one document. They are "was this invoice ever paid", "which payments have no invoice behind them", and "what did we spend with this vendor last quarter". Answering any of them today means opening files one at a time.
What it does
paperwork-reconciler indexes authorized NAS folders into a local ledger, extracting vendor, date, amount, document type and reference number from each document. It then matches payments to invoices on amount, a date window and reference number, and reports three lists:
- Matched — invoice and payment pairs, with a confidence score
- Unpaid — invoices with no payment found in the window
- Unmatched payments — payments and receipts with no invoice behind them
Every figure cites its source filename, so nothing has to be taken on trust. Two appendices cover possible matches below threshold and documents whose type or amount could not be resolved.
How it is different
Conventional document AI summarises one file at a time. This reconciles relationships between documents across years, and surfaces absence — the invoice with no payment, the payment with no invoice. Absence is what an owner actually needs at tax time, and no single-document tool can see it.
Why it belongs on a NAS
Not as a preference, but as the whole argument. These are invoices, bank statements and receipts. Extraction and inference run on the device, so no financial document ever leaves the building. The NAS holds the durable ledger between incremental runs, folder-watch triggers on arrival, and multi-user shares let a bookkeeper read the report without being given the source documents.
How we built it
Four expert-agent stages in order — enumerate, extract, reconcile, report — expressed in Markdown with no executable scripts. Extraction resolves five fixed fields and writes null with a reason rather than guessing. Vendor-name variants cluster by normalisation and then fuzzy match, with every merge recorded reversibly. Matching is deterministic and weighted: amount 0.50, date proximity 0.20, reference 0.20, payee 0.10, with outright rejection on a reference mismatch or a different payee.
Challenges
Deciding what not to do. Part payments, consolidated payments and credit notes are all real patterns in small-business paperwork, and each deserves proper treatment rather than a heuristic. They are reported as unresolved instead of quietly mismatched. The same thinking drove the contention rule: when two invoices match one payment above threshold, neither is matched. A plausible-looking wrong match is worse than an honest unresolved one — a user resolves an ambiguity in seconds but will not catch a confident error.
What's next
Binding each stage to the registered UGOS expert agents from the Uliya Skill Development Standard, then validating extraction quality on real scans and end-to-end run time over a large folder.
Built With
- local-ai
- markdown
- ugos
- ugreen-ai-nas